2026-06-11
First in a series distilled from months of living with long-memory agents before Throughline existed — every lesson here was paid for in real use. The patterns are public; they're in the product because we needed them, not because they demo well.
An investment copilot, ninety days in. It keeps a judgment ledger — every call logged with a conviction score and a falsifier ("what would prove me wrong"). One evening, its user is about to re-enter a position he'd been stopped out of twice, conviction high, typing fast.
Think about what a normal memory system does at this moment. It recalls what's most salient about that ticker — and salience grows with use. Which memories got used the most? The thesis he kept repeating. The bullish case he rehearsed across a dozen conversations. His own voice, played back as memory.
A memory that gets warmer with use is an echo chamber with receipts.
Every memory product on the market has one recall policy. Usually it's reinforcement: what you return to comes back faster — and for companionship that's exactly right; it's what makes a self feel alive instead of indexed. We use that loop too, gladly.
But run the same policy over judgments and you've built confirmation bias as infrastructure. You rehearse what you believe; the system reinforces what you rehearse; the memory most likely to surface at the moment of decision is the one that agrees with you. Nobody designed the flattery. The loop did.
The single most valuable mechanism in that copilot — by its user's own assessment — was the opposite of warmth. Before agreeing with any new high-conviction call, it was required to recall two things: every prior judgment on that ticker, and every falsified high-conviction judgment on record. Then it had to ask one question and hold it:
loading up again. this dip is different — the thesis is intact.
You were exactly this sure twice before — both entries are logged, both rolled back. What's structurally different this time? (Your conviction-4+ hit rate is 33%. I'll wait.)
Explicitly immune to impatience. A friend with receipts, not a mirror with a good memory.
So in Throughline, streams are typed by what the memory is for:
Relationship streams (shared history, callbacks, the texture of knowing someone) reinforce with use. What you return to stays on the tip of its tongue. Warmth is the product.
Governed streams (judgments, risks, lessons, failures) never reinforce. Recall there is lexical plus recency — deterministic, cold by design. No amount of rehearsing a thesis can promote it, and a regression test pins the property: salience must not reorder judgments. The ledger also computes your actual hit rate by conviction bucket and carries it into every session — calibration as a standing companion, not a postmortem.
One self, two temperatures. The part that comforts you deepens with use; the part that protects you refuses to.
Ask what each memory is for before deciding how it's allowed to behave. A single recall policy means choosing flattery for everything — which feels fine right up until the day it's expensive.
The advisor that remembers your calls · how the engine works · start free